← Tilbake til oversikten (norsk)

Unprocessed red meat and colorectal cancer incidence in prospective cohorts: an updated meta-analysis to October 2026 using the WCRF CUP 2016 method

Re-analysis series, article 6. Draft of 30 October 2026.

Abstract

Background. The WCRF Continuous Update Project (CUP) last pooled red meat and colorectal cancer in its 2016 systematic literature review: relative risk (RR) 1.12 (95 % CI 1.00–1.25) per 100 g/day from 8 cohort studies. At least a dozen cohorts have since published estimates for unprocessed red meat.

Methods. Pre-specified update (protocol rule R7). Europe PMC was searched from 1 May 2015 to 31 October 2026 (CUP outcome block plus meat terms, and a broad fallback query), with automated keyword triage followed by manual screening of 1,028 records. One row per cohort, closest to unprocessed red meat, was converted to RR per 100 g/day (Greenland–Longnecker for category data) and pooled by DerSimonian–Laird, with REML, Hartung–Knapp–Sidik–Jonkman (HKSJ) and fixed-effect sensitivity analyses, leave-one-out, subgroups, a 1,152-combination multiverse, a two-stage restricted cubic spline and negative controls (poultry, fish).

Results. 16,731 unique records yielded 22 eligible studies; of four new to our database, one entered a sensitivity analysis and three were excluded. The main pool had 20 cohorts and 45,117 cases. The summary RR was 1.10 (1.04–1.15), I² = 13 %; alternative estimators agreed. Every leave-one-out estimate and all 1,152 multiverse combinations had a lower limit above 1. There was no evidence of non-linearity (p = 0.54), but a linear slope on the 12 category sets used for the spline was steeper, 1.24 (1.06–1.44). Poultry gave 0.89 (0.74–1.08) and fish 0.76 (0.65–0.90).

Conclusions. With the literature to October 2026, the CUP 2016 method gives a slightly smaller point estimate than CUP 2016 with a much narrower interval, robust to single cohorts and extraction choices. The association is small, and the analysis does not address residual confounding or dietary measurement error, a point the inverse fish association underlines.

1. Introduction

The CUP 2016 dose–response estimate for red meat and colorectal cancer, 1.12 (1.00–1.25) per 100 g/day from 8 studies with I² = 24 %, was published in the peer-reviewed literature by Vieira et al. [1, p. 7]. It is the per-gram figure behind the CUP's colorectal judgement on red meat. The earlier articles in this series reproduced that estimate exactly from CUP's own per-study values. They then examined how it depends on extraction choices [2].

This article asks a narrower question: what does the CUP 2016 method give when it is applied to the cohort literature as it stands in October 2026? The CUP SLR search ended in 2015 [3]. Since then, cohorts in the United Kingdom (UK Biobank, Million Women Study), the United States (NIH-AARP, Sister Study, Black Women's Health Study), Canada, Denmark, Finland, Israel and Japan have published estimates for unprocessed red meat. EPIC, NHS/HPFS and the Netherlands Cohort Study have published updates. None of the three recent syntheses we read [1, 4, 5] pooled these with the CUP per-gram method.

We also report two things CUP did not: a negative-control analysis with poultry and fish from the same database, and a comparison between the linear slope estimated from category data and the pooled slope that is dominated by published continuous estimates.

2. Methods

2.1 Protocol

The analysis follows rule R7 of our re-analysis protocol, written before the search was run [6]. R7 specifies a Europe PMC search with the CUP SLR terms (1 May 2015–31 October 2026), a screening log, inclusion by the SLR 2016 rules (prospective design, incidence, red meat separable from processed meat where possible, most recent or most complete publication per cohort), DerSimonian–Laird pooling per 100 g/day with HKSJ as sensitivity, a restricted cubic spline where at least five studies have three or more categories, subgroups by sex and region, leave-one-out, Egger's test and a multiverse over extraction choices.

Everything below that goes beyond R7 is marked as additional: the REML and fixed-effect pools, the negative controls, the linear slope on the spline sets, and the subgroups by publication period and study size.

The search was run on 30 October 2026. The full log, including the exact query strings, is published with the data [7]. There were two queries.

Query (a). The CUP outcome block, translated into Europe PMC syntax. [MeSH] became MESH: OR KW:, [tiab] became TITLE_ABS:, and colon and rectal neoplasms were listed explicitly. The outcome block was combined with meat terms only (meat, beef, pork, lamb, processed, cured, "red meat") and the date window. CUP had instead combined the outcome block with its whole nutrition block and screened by hand. Our narrower combination is a deliberate restriction, and a deviation from R7's "CUP search terms"; the broad query (b) and the sensitivity check below were the safeguards.

Query (b). A broad fallback: ("red meat" OR "unprocessed meat" OR beef OR pork OR lamb) AND (colorectal OR colon OR rectal OR rectum) AND (cohort OR prospective OR "case-cohort" OR "nested"), with the same window. Unfielded Europe PMC terms also search PMC full texts, so this query is large and noisy.

Other sources. We checked the reference lists of the three recent meta-analyses [1, 4, 5], and ran three targeted probes by cohort name (for example, Kadoorie, Malmö Diet, Diet Cancer and Health, NutriNet).

Screening. With over 16,000 unique records, screening had two stages.

Stage 1 was automated keyword triage on title and abstract. A record went forward only if it contained all three of: - a colorectal term; - a meat term or a general diet term (diet terms were allowed so that diet-wide and risk-factor papers would not be lost); - a prospective-design term.

Records without an abstract went forward if the title had a colorectal or meat term. As a check, all 155 stage-1 exclusions whose title contained both a meat word and a cancer word were read by hand. Four were pulled back for abstract screening; none qualified.

Stage 2 was manual title screening of every record that passed stage 1. Abstracts were then read for every record not clearly ineligible from its title, and full texts were checked for three records.

Sensitivity check. We tested whether the search retrieved every cohort publication from 2015 onward already in our database.

2.3 Eligibility and choice of estimate

We included prospective cohort, case-cohort and nested case-control studies of adults that reported colorectal cancer incidence by red meat intake, with an estimate convertible to grams per day. We kept one row per cohort. Cohort overlap was resolved with the parent–child table in the database [8]: for example, Malmö Diet and Cancer is a centre of EPIC, and JPHC-5y overlaps JPHC I+II.

Within each cohort we took the main multivariable model and the exposure closest to unprocessed red meat. When a later publication superseded an earlier one, the later one entered the main pool. The earlier one entered the multiverse as an alternative where the two differ in a substantive choice. Examples are EPIC (Papadimitriou 2022 vs Norat 2005 observed and calibrated), UK Biobank (Knuppel 2020 vs Bradbury 2020), and JPHC/Takayama (Islam 2019 vs Wada 2017).

For the cohorts CUP 2016 pooled for colorectal cancer, six entered through the same publication (ATBC, Dutch Monitoring Project, MCCS, SMC, MEC, SWHS). For MEC we used the paper's unprocessed row, and for the Monitoring Project our own Greenland–Longnecker value, rather than CUP's rows. EPIC was updated to Papadimitriou 2022. Järvinen 2001 was excluded because its exposure is total meat. Section 3.1 lists every exclusion and the reason for it.

2.4 Conversion to a common scale

Each row was put on the scale "RR per 100 g/day". The route depended on how the paper reported its result.

2.5 Statistical analysis

Software. All runs come from one script, scripts/update2026.py, which writes kontroll/reanalyser/update2026.md and .json. Pooling uses scripts/pool.py. The Greenland–Longnecker routine is scripts/gl_doseresponse.py. The spline basis, the Greenland–Longnecker covariance, multivariate REML, univariate REML, HKSJ and Egger's test are in scripts/mvmeta.py.

2.6 Extraction and control

The database [8] was built and checked by AI agents (Anthropic Claude Opus instances) working under the author's direction to a fixed codebook. It holds 10,383 estimates from 53 publications. Every row was checked against the full text or supplement by instances that had not seen the extraction: - 10,353 rows were verified; - 18 were corrected; - 12 are disputed because the source paper's own numbers are internally impossible.

Flag, unit and metadata corrections are listed in a correction log [16]. The Vulcan 2017 rows were added for this update.

The search screening (stage 2) was also done by an AI agent, and the stage-1 triage was automated. As protocol §6 requires, the pooled numbers were recomputed independently. The 46 re-analysis runs of the earlier syntheses were checked with a second script (kontroll/meta_check.py, written by the K11 controller without sight of pool.py; output regenerated in kontroll/meta_check_reanalyser.txt): 45 agreed to four decimals (K14), and the remaining one, WCRF 2007 Fig 4.3.3 per 100 g, was confirmed exactly in K16a. The 2026 analyses were recomputed separately (kontroll/K16a/meta_check2.py, kontroll/K16a/RAPPORT.md). The main pool, all subgroups, the sensitivity runs, leave-one-out, Egger's test, HKSJ, fixed effect and REML agreed to four decimals; the main pool is 1.0967 (1.0447–1.1513) in both. The multiverse result (1,152 of 1,152 combinations with a lower limit above 1; median and range) was reproduced. The negative controls and the Greenland–Longnecker slopes could be reproduced only approximately, because their per-study inputs are not stored in the results file: fish matched to two decimals, poultry gave 0.90 against our 0.89, and the linear slope on the 12 spline sets gave 1.23 against our 1.24. These three estimates should therefore be read as reproducible to about ±0.01, not exactly.

3. Results

3.1 Search and selection

Table 1 gives the counts. After 630 duplicates were removed, 16,731 records were screened. Automated triage excluded 15,703 of them, mostly for lacking a colorectal term (11,541). Manual screening of the remaining 1,028 left 22 eligible studies, and the four records added from reference lists and probes were all excluded.

The sensitivity check retrieved all 18 cohort publications from 2015 onward that were already in the database. Two of them (Hoang 2023 and Jin 2023) were found only by query (b), and they passed stage 1 only because of the diet-term allowance. Query (a) on its own therefore had a sensitivity of 16/18 against the database. All three recent meta-analyses [1, 4, 5] were also retrieved.

Table 1. Search and selection (PRISMA counts)

Step n
Records identified, query (a) 8,119
Records identified, query (b) 9,242
Additional records: Ungvari 2025 reference list / cohort-name probes 2 / 2
Duplicates removed (520 in both queries; 110 within-query) 630
Records screened (search) 16,731
Excluded at stage 1 (automated keyword triage) 15,703
Screened manually at stage 2 (title, then abstract; full text for 3) 1,028
Excluded at stage 2: design / outcome / exposure / duplicate / other 478 / 261 / 256 / 6 / 5
Additional records excluded (exposure 3, duplicate 1) 4
Eligible studies 22
– already in the database 18
– new 4
Cohorts in the main pool 20

Of the 18 database publications from 2015 onward, 13 enter the main pool. Wada 2017 is replaced by the pooled Islam 2019. Etemadi 2018 is an earlier NIH-AARP analysis. Bradbury 2020 and Hoang 2023 are UK Biobank papers replaced by Knuppel 2020. Jin 2023 reports results per SD with no SD in grams. The remaining seven cohorts in the pool were published before 2015.

The four new studies:

All four borderline exclusions flagged in the search log are now resolved; the author obtained the closed-access full texts.

The PRISMA counts in Table 1 are those of the original screening log.

Exclusions of other database studies, with reasons, are listed in update2026.md.

3.2 Study characteristics

Table 2. Cohorts in the main pool (one row per cohort). RR per 100 g/day of red meat, as entered into the pool

Cohort (publication) Region Sex Exposure as defined Cases RR (95 % CI) Weight % Route to per 100 g/day
ATBC (Pietinen 1999) [17] Europe men beef, pork, lamb 185 0.83 (0.43–1.61) 0.5 GL, quartile medians
Dutch Monitoring Project (Tiemersma 2002) [18] Europe both fresh red meat 102 1.92 (0.85–4.31) 0.4 GL, CUP dose rule; nested case-control
MCCS (English 2004) [19] Australia both fresh red meat 451 1.19 (0.89–1.57) 2.8 per serving/week × 5.833 (120 g)
EPIC (Papadimitriou 2022) [20] Europe both red meat (not defined in paper; EPIC convention fresh) 5,069 1.06 (0.95–1.14) 16.3 per SD (36.2 g)^2.762
SMC (Larsson 2005) [21] Europe women beef/pork main dish 733 1.23 (0.90–1.67) 2.3 GL, category medians × 120 g
MEC (Ollberding 2012) [22] N America both red meat excluding processed 3,404 0.97 (0.83–1.13) 8.2 GL, SLR energy conversion
NLCS (Gilsing 2015) [23] Europe both fresh red meat 437 1.00 (0.81–1.23) 4.8 per 50 g^2; NLCS Meat Investigation Cohort, full-cohort analysis
JPHC + Takayama (Islam 2019) [24] Asia both unprocessed red meat 3,550 1.09 (0.91–1.31) 6.1 per 100 g; sexes combined (fixed)
UK Biobank (Knuppel 2020) [25] Europe both unprocessed red meat 3,201 1.50 (1.10–2.04) 2.3 GL, category means, model 2
DANSDA (Mejborn 2021) [26] Europe both unprocessed red meat 127 1.04 (0.67–1.61) 1.2 per 100 g
Israel, three cohorts (Dankner 2024) [27] Other both unprocessed red meat 813 1.15 (0.87–1.52) 2.8 per 100 g (authors' pooled)
Sister Study (Mehta 2020) [28] N America women unprocessed red meat 210 1.30 (0.77–2.19) 0.9 per 50 g^2
NIH-AARP (Zouiouich 2025) [29] N America both unprocessed red meat (coded partial) 9,461 1.13 (1.06–1.20) 25.6 per 50 g/1,000 kcal ≈ 100 g/day at 2,000 kcal (assumption)
NHS + HPFS (Bernstein 2015) [30] N America both unprocessed red meat incl. hamburger 2,731 1.00 (0.89–1.11) 13.4 per 100 g, Supplement S5
Million Women (Papier 2025) [31] Europe women red meat (unprocessed, partial) 12,251 1.28 (1.00–1.54) 4.6 per 20 g^5, measurement-error calibrated
Alberta's Tomorrow (Al Rajabi 2021) [32] N America both unprocessed red meat 257 1.25 (0.68–2.28) 0.6 GL, CUP dose rule; sexes combined
BWHS (Yiannakou 2022) [33] N America women unprocessed red meat (partial) 564 1.63 (0.84–3.14) 0.5 GL from tertiles
Finnish cohorts (Tammi 2024) [34] Europe both unprocessed red meat 1,124 1.23 (1.00–1.50) 5.0 per 100 g/week^7
Korea (Wie 2014) [35] Asia both unprocessed red meat 53 0.90 (0.48–1.79) 0.5 per 10 g^10
SWHS (Lee 2009) [36] Asia women red meat (undefined; mainly pork) 394 0.81 (0.50–1.29) 1.0 CUP 2016's extracted row
Total 45,117 100

GL: Greenland–Longnecker. Cases are those in the analysis that produced the estimate used (Dankner 2024: 813 of 839; Al Rajabi 2021: 257 in Model 2 of 265). Database estimate identifiers for each row are listed in kontroll/reanalyser/update2026.md.

Weights. Three cohorts carry 55 % of the random-effects weight, and each of them enters as a published continuous estimate or a direct rescaling of one: NIH-AARP (25.6 %), EPIC (16.3 %) and NHS/HPFS (13.4 %). MEC adds 8.2 %.

Rescaling and rounding. Several rows are rescaled from small increments, so they carry the rounding of the published confidence interval. The EPIC row is the clearest case. Papadimitriou's published 1.02 (0.98–1.05) per SD becomes visibly asymmetric, 1.06 (0.95–1.14), after it is raised to the power 2.762.

3.3 Main result

The summary RR per 100 g/day was 1.10 (1.04–1.15) (run: update2026 – main, DerSimonian–Laird), from k = 20 cohorts, with I² = 13 %, τ² = 0.0014 and Q = 21.8 on 19 df. The alternative estimators agree (Table 3). Egger's test gave no evidence of small-study asymmetry (intercept 0.33, p = 0.40).

Table 3. Pooled estimates, RR per 100 g/day (all runs from update2026.json)

Run k RR (95 % CI) I²
CUP 2016 as published (Vieira 2017 [1, p. 7]; reproduced exactly [2]) 8 1.12 (1.00–1.25) 24 %
Main: DerSimonian–Laird 20 1.10 (1.04–1.15) 13 %
REML (τ² 0.0016) 20 1.10 (1.04–1.15) –
HKSJ 20 1.10 (1.04–1.16) –
Fixed effect 20 1.10 (1.06–1.14) –
Leave-one-out, lowest 19 1.09 (1.03–1.15) (without NIH-AARP) –
Leave-one-out, highest 19 1.11 (1.06–1.16) (without NHS + HPFS) –
Without Million Women (calibrated) 19 1.09 (1.04–1.14) 9 %
Without NIH-AARP (energy-density assumption) 19 1.09 (1.03–1.15) 12 %
Without nested case-control (Tiemersma) 19 1.09 (1.04–1.15) 10 %
Without the two largest (MWS, NIH-AARP) 18 1.07 (1.01–1.13) 5 %
With Malmö DCS (Vulcan 2017) despite EPIC overlap 21 1.09 (1.05–1.15) 9 %
NHS/HPFS → Wang 2020 (NHS + NHSII + HPFS; per 3 servings/week → per 100 g at 120 g/serving; age strata combined) 20 1.10 (1.06–1.15) 3 %
Multiverse, 1,152 combinations: median (range) 20 1.11 (1.09–1.13); lower limit > 1 in 1,152/1,152 –
Spline (12 category sets, 10 cohorts): RR at 100 g vs 0 12 sets 1.23 (1.06–1.44) –
Linear two-stage slope, same 12 sets 12 sets 1.24 (1.06–1.44) –
Negative control: poultry 7 0.89 (0.74–1.08) 40 %
Negative control: fish 5 0.76 (0.65–0.90) 0 %

Figure 1 (placeholder): forest plot of the 20 cohorts with the random-effects diamond, and the spline curve with its 95 % band, to be generated from update2026.json in the HTML companion.

3.4 Leave-one-out

All 20 leave-one-out estimates fall between 1.09 and 1.11, and the lowest lower limit is 1.03, reached when NIH-AARP, the heaviest cohort, is removed (Table 3). Removing NHS/HPFS, MEC or EPIC raises the estimate to 1.11.

3.5 Subgroups

Table 4. Subgroups (DerSimonian–Laird)

Subgroup k RR (95 % CI) I²
Europe 9 1.15 (1.04–1.27) 27 %
North America 6 1.07 (0.98–1.17) 38 %
Asia 3 1.04 (0.88–1.22) 0 %
Women-only cohorts 5 1.22 (1.05–1.43) 0 %
Mixed or men-only cohorts 15 1.08 (1.03–1.14) 12 %
Explicitly unprocessed (database flag = no; excludes partial/undefined rows) 14 1.11 (1.03–1.19) 0 %
Partial or undefined exposure rows only (EPIC 2022, NIH-AARP, NHS + HPFS, MWS, BWHS, SWHS) 6 1.09 (1.01–1.18) 45 %
Published 2015 or later 13 1.11 (1.05–1.17) 13 %
Published before 2015 7 1.04 (0.91–1.19) 10 %
≥ 1,000 cases 8 1.10 (1.03–1.18) 49 %
< 1,000 cases 12 1.11 (0.99–1.23) 0 %

The Australian (MCCS) and Israeli cohorts fall outside the three regional subgroups.

The exposure subgroups are defined by the database flag for processed meat. The 14 cohorts flagged as unprocessed give 1.11 (1.03–1.19) with no heterogeneity. The six rows flagged partial or undefined give 1.09 (1.01–1.18) with I² = 45 %. These are EPIC 2022, NIH-AARP, NHS + HPFS (which includes hamburger), Million Women, BWHS and SWHS. Restricting the pool to cleanly defined exposures therefore does not lower the estimate.

The subgroup differences are modest, and the intervals overlap. No test of interaction was pre-specified, and we draw no conclusion from them.

3.6 Multiverse

All 1,152 combinations of this update's multiverse (not to be confused with the 1,152 = 768 + 384 combinations of the CUP 2016 multiverse) had a lower confidence limit above 1, with a median RR of 1.11 and a range of 1.09–1.13. No single choice moved the median by more than 0.01. The alternatives covered the Dutch Monitoring Project (CUP's value), MCCS (adjusted for processed meat), EPIC (Norat 2005 observed or calibrated), SMC and MEC (total red meat), JPHC/Takayama (Takayama alone), UK Biobank (Bradbury 2020), NIH-AARP (1,800 or 2,200 kcal) and the model (random or fixed).

This contrasts with the multiverse built around CUP 2016's own eight studies. There, only 16 % of 768 combinations had a lower limit above 1, and the EPIC choice carried most of that [2, 37]. Adding the post-2015 cohorts removes the dependence of statistical significance on any one of the extraction choices tested.

3.7 Non-linear dose–response

Twelve category sets from ten cohorts met the requirements for the spline: ATBC, the Dutch Monitoring Project, SMC, EPIC (Norat 2005), MEC, UK Biobank (Knuppel 2020), Sister Study, Takayama (Wada 2017, men and women), Alberta's Tomorrow (men and women) and BWHS. Knots fell at 9, 38 and 84 g/day. For SMC and Alberta's Tomorrow the case distribution across categories was not printed and had to be assumed.

There was no evidence of departure from linearity (Wald χ² = 0.38, df = 1, p = 0.54). Residual heterogeneity was low (Q = 20.0 on 22 df, p = 0.58). Relative to zero intake, the fitted curve gave 1.07 (1.00–1.16) at 25 g/day, 1.14 (1.01–1.28) at 50, 1.19 (1.04–1.35) at 75, 1.23 (1.06–1.44) at 100, 1.28 (1.05–1.55) at 125 and 1.32 (1.04–1.69) at 150 g/day.

A linear Greenland–Longnecker slope fitted to the same 12 sets and pooled by DerSimonian–Laird gave 1.24 (1.06–1.44) per 100 g/day. The spline curve and the linear slope on the same sets therefore agree. Both are steeper than the main pool.

3.8 Negative controls

Poultry, pooled across seven cohorts with the same method (Dankner 2024, English 2004, Gilsing 2015, Mejborn 2021, Norat 2005, Spencer 2010, Bradbury 2020), gave 0.89 (0.74–1.08) with I² = 40 %. Fish, from five cohorts (English 2004, Gilsing 2015, Norat 2005, Spencer 2010, Bradbury 2020), gave 0.76 (0.65–0.90) with I² = 0 %.

These pools use whichever publication of each cohort reports poultry or fish. For EPIC that is Norat 2005, and for UK Biobank it is Bradbury 2020. They also include Spencer 2010, which is excluded from the red-meat pool because it overlaps EPIC. The negative-control set is therefore not the same set of cohorts as the red-meat pool, and only English 2004, Gilsing 2015, Mejborn 2021 and Dankner 2024 contribute to both pools in the same publication.

4. Discussion

4.1 The update compared with CUP 2016

With 20 cohorts and 45,117 cases, the CUP 2016 method gives 1.10 (1.04–1.15) per 100 g/day. CUP 2016 had 1.12 (1.00–1.25) from 8 studies [1, p. 7]. The point estimates hardly differ, but the interval is less than half as wide. In the update, no estimator and no single omission moves the lower limit below 1.03, and every multiverse combination keeps it above 1.

This matters given what the earlier articles in this series found. In CUP 2016's own study set, significance depended on a handful of defensible choices. With CUP's extraction rules applied consistently, the colorectal estimate was 1.08 (0.93–1.26) [2, 37]. The update suggests that this fragility reflected the small 2016 evidence base: with the post-2015 cohorts the pooled estimate is stable at about 1.10. The post-2015 cohorts are large and mostly report unprocessed red meat directly.

Two cautions follow. First, stability is not the same as validity. A precise pooled estimate from observational studies that share the same sources of bias is precisely biased if the bias is there. Section 4.4 returns to this.

Second, the effect size is small. An RR of 1.10 per 100 g/day corresponds to roughly a 5 % relative difference at 50 g/day (linear interpolation). The spline gives 1.14 at 50 g/day relative to zero intake, from a different set of studies (Section 4.2).

4.2 Category-based slopes are steeper than the continuous-estimate pool

The spline and the linear two-stage slope on the 12 category sets agree with each other at about 1.23–1.24 per 100 g/day, and both are steeper than the main pool's 1.10. This is a methodological observation, not evidence of a stronger causal effect. At least four differences between the two analyses could produce it, and our runs do not separate them.

  1. Different cohorts. The 12 sets do not include the three cohorts that carry 55 % of the main-pool weight (NIH-AARP, EPIC 2022 and NHS/HPFS), nor Million Women, NLCS, the Finnish cohorts, Israel or DANSDA. All of those report continuous estimates and no usable category sets. The spline sets use Norat 2005 for EPIC and Wada 2017 for Takayama, not the later publications in the main pool.
  2. Dose assignment. Category-based slopes depend on the dose given to each category: medians, midpoints, 120 g per serving, or energy-based conversion. If the assigned doses compress the true contrast between categories, the slope per gram is inflated. If they stretch it, the slope is attenuated. The direction of this error is unknown for most studies.
  3. The published continuous models of the large cohorts have their own modelling choices. Examples are the NIH-AARP addition model per 1,000 kcal and the EPIC per-SD model rescaled beyond one SD. The Greenland–Longnecker reconstruction does not share those choices.
  4. Small-study composition. Within the main pool, the cohorts with fewer than 1,000 cases give 1.11 (0.99–1.23) and the larger ones 1.10 (1.03–1.18), so study size explains little of the gap. Egger's test in the main pool shows no asymmetry.

The practical point is that the choice between a category-based and a continuous-estimate-based pool can shift the slope by more than the difference between CUP 2016 and this update. It would be wrong to read the spline as showing that the "true" slope is 1.24 and the main pool as diluted, or the reverse. They answer slightly different questions on different subsets of the literature.

4.3 The negative controls

Poultry is the more natural negative control: a meat eaten in similar settings, without the haem and processing-related exposures proposed as mechanisms for red meat. It gave 0.89 (0.74–1.08). The interval includes 1, and the point estimate is below 1.

Fish gave 0.76 (0.65–0.90), a clearly inverse association in five cohorts. Vieira 2017 reported 0.89 (0.80–0.99) per 100 g/day for fish from 11 studies [1, p. 9] and 0.81 (0.53–1.25) for poultry from 7 studies [1, p. 10]. Our fish estimate is therefore more inverse than CUP's, from fewer cohorts.

A clean negative control should be null. Fish is either protective, or it marks a pattern of diet and health behaviour, or both. In either case it is not a clean negative control. In cohorts that overlap with the red-meat pool, with the same kind of questionnaires and adjustment sets, a food that public-health advice favours shows an inverse association whose absolute size on the log scale is roughly three times that of the red meat association. Food-level associations of this size can arise from the dietary pattern a food stands for. The red meat estimate is open to the same explanation in the opposite direction, and these data cannot rule that out. The control does not show that the red meat association is confounded. It shows that the method is capable of producing associations of comparable or larger size for foods where confounding by healthy-eater behaviour is plausible.

4.4 Limitations

Energy-density assumption for NIH-AARP. The heaviest cohort (25.6 % of the weight) enters through the assumption that 50 g/1,000 kcal equals 100 g/day at 2,000 kcal/day. Taking 1,800 or 2,200 kcal moves the multiverse median by at most 0.01. Removing NIH-AARP gives 1.09 (1.03–1.15). The assumption ignores how energy intake is distributed within the cohort, so the row is an approximation, though the result does not depend on it.

The calibrated Million Women row. Papier 2025 prints red meat alone only as a measurement-error-corrected (calibrated) estimate in Supplementary Data 1. We used it as a documented exception to our rule of using uncalibrated rows (correction log R22 [16]). Removing the row gives 1.09 (1.04–1.14). The MWS exposure is also coded "partial" for processing.

Exposure definitions. Several "unprocessed" variables are coded partial in the database. NHS/HPFS includes hamburger, and NIH-AARP and BWHS are partially separable. EPIC's paper does not define red meat, and SWHS is undefined and mainly pork. Restricting to the 14 cohorts flagged unprocessed gives 1.11 (1.03–1.19) (Section 3.5).

Automated stage-1 screening. 15,703 records were excluded by keyword rules, not by reading. One class of paper is at risk: a cohort study that reports red meat only in its tables and has no colorectal, diet or design term in its abstract. Hoang 2023 and Jin 2023 are of this type and were caught only because diet terms were allowed. The title check and probes reduce this risk. Embase and Web of Science were not searched. The closed-access borderline papers were obtained and checked after screening. Only Wang 2020 has a usable estimate, and it needs an assumed serving size; replacing NHS/HPFS (Bernstein 2015) with Wang 2020 gives 1.10 (1.06–1.15) with I² = 3 %.

AI-agent extraction and screening. Extraction, control, stage-2 screening and the analyses were all carried out by AI agents under the author's direction. The independent row-level control (10,353 of 10,383 rows verified; 18 corrected, 12 disputed) and the correction log are the safeguards, and every row's provenance is printed. The independent recomputation (K16a) reproduced the main pool, subgroups, sensitivity runs, leave-one-out, estimators and multiverse; the negative controls and the Greenland–Longnecker slopes (linear slope and spline sets) matched only to about two decimals, because their per-study inputs are not stored in the results file (Section 2.6).

Residual confounding and measurement error. This analysis addresses neither. Food-frequency measurement error, residual confounding by correlated diet and lifestyle factors, and differences in adjustment sets across cohorts are all inherited unchanged. The analysis shows how stable the CUP-method estimate is, not whether the association is causal.

Control K19 (kontroll/K19/) documented that the dose–response code rescales the Greenland–Longnecker between-category correlation to the published standard errors rather than using the textbook GL/Orsini covariance. All per-study slopes reproduce to four decimals under that documented rule. With textbook GL the main pool is 1.0955 (1.0433–1.1503) instead of 1.0967 (1.0447–1.1513), the linear two-stage slope 1.23 (1.05–1.43) instead of 1.24 (1.06–1.44), and single-study slopes move by up to ±0.06 (SMC 1.23 → 1.17). No conclusion changes; the variant is now stated in the protocol change log.

4.5 Comparison with other recent syntheses

Three syntheses published since CUP 2016 report pooled estimates on different scales, so only their stated values are compared here.

Vieira et al. (2017) is the peer-reviewed publication of the CUP 2016 analysis. It reports 1.12 (1.00–1.25) per 100 g/day for colorectal cancer, with I² = 24 % and 8 studies, 1.22 (1.06–1.39) for colon cancer and 1.13 (0.96–1.34) for rectal cancer [1, p. 7]. Our update is the same analysis carried forward to October 2026.

Han et al. (2019) expressed results as the RR for a reduction of 3 servings of unprocessed red meat per week. For colorectal cancer incidence they report 1.00 (0.92–1.09) from 5 studies and 322,502 participants, rated low-certainty evidence [4, Table 1, p. 714]. The scale (servings, reduction) and the study set differ from ours, and their per-study inputs are in a supplement we do not hold. Their null estimate and our 1.10 per 100 g/day are therefore not directly comparable, and the difference cannot be resolved from the files available here.

Ungvari et al. (2025) pooled highest vs lowest intake from 60 prospective studies. For red meat and colorectal cancer they report a hazard ratio of 1.15 (1.10–1.21) [5, abstract, p. 5123]. Across colon, colorectal and rectal outcomes combined it is 1.18 (1.13–1.23), from 71 cohort entries with I² = 29 % [5, p. 5127]. This is a category contrast, not a dose–response slope. Their red meat "generally referred to beef, pork, lamb, mutton, and veal" [5, p. 5125], so it was not restricted to unprocessed meat in every study. Their larger point estimate is consistent in direction with ours. The magnitudes are not comparable without the intake difference behind each highest-vs-lowest contrast.

5. Conclusion

The CUP 2016 method, applied to the prospective cohort literature on unprocessed red meat and colorectal cancer to October 2026, gives an RR of 1.10 (1.04–1.15) per 100 g/day. It shows little heterogeneity and is insensitive to estimator, single cohorts and extraction choices. The 2016 estimate depended on such choices; this one does not. Category-based slopes on a smaller, different subset are steeper, at about 1.24. In overlapping cohorts, fish shows an inverse association larger than the red meat association. Both observations limit how far the stable pooled number can be read as a causal effect size.

Data and code availability

The raw Europe PMC downloads (about 170 MB) were not kept and can be regenerated with kontroll/sok2026/scripts/fetch.py.

References

  1. Vieira AR, Abar L, Chan DSM, Vingeliene S, Polemiti E, Stevens C, Greenwood D, Norat T. Foods and beverages and colorectal cancer risk: a systematic review and meta-analysis of cohort studies, an update of the evidence of the WCRF-AICR Continuous Update Project. Ann Oncol 2017. PMID 28407090. (Page numbers refer to the accepted manuscript.)
  2. Sunnanå LM. CUP 2016 re-analysed: red meat and colorectal cancer. Re-analysis series (companion article, artikler/cup2016.md), and database summary database-samledom.md, Tests 1–4.
  3. World Cancer Research Fund / American Institute for Cancer Research. Continuous Update Project systematic literature review: colorectal cancer, 2016.
  4. Han MA, Zeraatkar D, Guyatt GH, et al. Reduction of red and processed meat intake and cancer mortality and incidence: a systematic review and meta-analysis of cohort studies. Ann Intern Med 2019;171:711–720. doi:10.7326/M19-0699.
  5. Ungvari Z, Fekete M, Varga P, et al. Association between red and processed meat consumption and colorectal cancer risk: a comprehensive meta-analysis of prospective studies. GeroScience 2025;47:5123–5140. doi:10.1007/s11357-025-01646-1.
  6. Re-analysis protocol, rules R1–R7 (REANALYSE_PROTOKOLL.md).
  7. Search log for the 2026 update (kontroll/sok2026/SEARCH_LOG.md), 30 October 2026.
  8. Red meat and colorectal cancer database, version 3, 30 October 2026 (database-samledom.md).
  9. Greenland S, Longnecker MP. Methods for trend estimation from summarized dose-response data, with applications to meta-analysis. Am J Epidemiol 1992;135:1301–1309.
  10. DerSimonian R, Laird N. Meta-analysis in clinical trials. Control Clin Trials 1986;7:177–188.
  11. Hartung J, Knapp G. A refined method for the meta-analysis of controlled clinical trials with binary outcome. Stat Med 2001;20:3875–3889.
  12. Sidik K, Jonkman JN. A simple confidence interval for meta-analysis. Stat Med 2002;21:3153–3159.
  13. Egger M, Davey Smith G, Schneider M, Minder C. Bias in meta-analysis detected by a simple, graphical test. BMJ 1997;315:629–634.
  14. Orsini N, Bellocco R, Greenland S. Generalized least squares for trend estimation of summarized dose-response data. Stata J 2006;6:40–57.
  15. Orsini N, Li R, Wolk A, Khudyakov P, Spiegelman D. Meta-analysis for linear and nonlinear dose-response relations: examples, an evaluation of approximations, and software. Am J Epidemiol 2012;175:66–73.
  16. Correction log (kontroll/rettelogg.md).
  17. Pietinen P, Malila N, Virtanen M, et al. Diet and risk of colorectal cancer in a cohort of Finnish men. Cancer Causes Control 1999;10:387–396.
  18. Tiemersma EW, et al. Cancer Causes Control 2002. doi:10.1023/a:1015236701054. PMID 12074508.
  19. English DR, MacInnis RJ, Hodge AM, et al. Red meat, chicken, and fish consumption and risk of colorectal cancer. Cancer Epidemiol Biomarkers Prev 2004;13:1509–1514. PMID 15342453.
  20. Papadimitriou N, et al. Clin Gastroenterol Hepatol 2022. doi:10.1016/j.cgh.2021.04.028.
  21. Larsson SC, Rafter J, Holmberg L, et al. Red meat consumption and risk of cancers of the proximal colon, distal colon and rectum: the Swedish Mammography Cohort. Int J Cancer 2005. doi:10.1002/ijc.20658.
  22. Ollberding NJ, et al. Int J Cancer 2012. doi:10.1002/ijc.27546.
  23. Gilsing AMJ, et al. Sci Rep 2015. doi:10.1038/srep13484. PMID 26316135.
  24. Islam Z, et al. Meat subtypes and colorectal cancer risk: a pooled analysis of 6 cohort studies in Japan. 2019. PMID 31482651.
  25. Knuppel A, et al. Int J Epidemiol 2020. doi:10.1093/ije/dyaa142.
  26. Mejborn H, et al. Nutrients 2021. doi:10.3390/nu13010032. PMID 33374887.
  27. Dankner R, et al. Int J Cancer 2024. doi:10.1002/ijc.35084.
  28. Mehta SS, et al. Cancer Epidemiol Biomarkers Prev 2020. doi:10.1158/1055-9965.EPI-19-0459. PMID 31575555.
  29. Zouiouich S, et al. Curr Dev Nutr 2025. doi:10.1016/j.cdnut.2025.107540.
  30. Bernstein AM, et al. PLoS ONE 2015. doi:10.1371/journal.pone.0135959.
  31. Papier K, et al. Nat Commun 2025. doi:10.1038/s41467-024-55219-5.
  32. Al Rajabi A, et al. Br J Nutr 2021. doi:10.1017/S0007114521001240.
  33. Yiannakou I, et al. J Nutr 2022. doi:10.1093/jn/nxab419.
  34. Tammi R, et al. Eur J Epidemiol 2024. doi:10.1007/s10654-024-01096-7. PMID 38253935.
  35. Wie GA, Cho YA, Kang HH, et al. Red meat consumption is associated with an increased overall cancer risk: a prospective cohort study in Korea. Br J Nutr 2014. doi:10.1017/S0007114514000683.
  36. Lee SA, et al. Nutr Cancer 2009. doi:10.1080/01635580802419780. PMID 19235035.
  37. Multiverse report for CUP 2016 (kontroll/multiverse_report.md) and scenarios report (kontroll/scenarios_report.md).

Studies used only in sensitivity, multiverse or negative-control runs: Norat T, et al. J Natl Cancer Inst 2005 (doi:10.1093/jnci/dji164); Wada K, et al. Cancer Sci 2017 (PMID 28256076); Bradbury KE, et al. Int J Epidemiol 2020 (doi:10.1093/ije/dyz064); Spencer EA, et al. Cancer Causes Control 2010 (doi:10.1007/s10552-010-9569-7); Vulcan A, Manjer J, Ericson U, Ohlsson B. Food Nutr Res 2017;61:1341810 (doi:10.1080/16546628.2017.1341810). Wang 2020, Gastroenterology (NHS, NHSII, HPFS; PMID 32247020). Checked and excluded after screening: Beslay M, et al. Eur J Nutr 2021;60:1887–96 (doi:10.1007/s00394-020-02381-3); Nguyen 2020, Int J Epidemiol (PMID 31203367).

Generated from the open database and scripts described in the article; extraction and control performed by AI agents (Claude Opus/Fable) under the author's direction. 30 October 2026.